Artificial intelligence is being deployed in medicine, engineering, finance, and other high-stakes professional settings at an accelerating pace. The assumption behind each deployment is straightforward: if a model scores well on its benchmarks, it is ready for the real world.
Mayank Ravishankara, a Bengaluru‑born software engineer and independent AI researcher, doesn’t fully agree.
Over the past three years, the Carnegie Mellon‑trained software engineer has quietly built a body of work that questions how the field measures understanding itself. Across papers, peer reviews, and hackathon judging, Ravishankara has been probing whether today’s AI systems truly comprehend the tasks they perform, or just recognize patterns that let them appear competent.
“Most evaluation benchmarks tell you whether a model got the right answer,” Ravishankara says. “They don’t tell you whether they actually understood the problem. In medicine, engineering, or finance, that gap carries real consequences.”
The Compliance Competence Trade‑Off
Through two papers presented at IEEE SoutheastCon 2026 (CircuChain and PlotChain) Ravishankara demonstrated a hidden flaw in conventional model testing. His experiments showed that powerful AI models can deliver seemingly correct answers simply by identifying surface‑level cues in benchmark formats rather than truly reasoning through the underlying task. When the same question is posed in a slightly different form, performance collapses.
In CircuChain, Ravishankara introduced the “Compliance-Competence Trade‑Off” to describe the difference between a system following instructions correctly (compliance) and actually understanding the problem (competence). His findings revealed something unsettling: two models with identical accuracy scores could be failing for entirely different reasons. One failed because it misunderstood the question, the other because it disobeyed it, a distinction that current benchmarks rarely expose.
Beyond Accuracy Scores
That insight extended into Ravishankara’s broader survey paper, “The Artificial Intelligence Cognitive Examination,” published in IEEE Access, one of the world’s largest open‑access engineering journals. The paper was subsequently featured by The Science Matters, an independent science journalism platform. To date, Ravishankara has published three peer‑reviewed papers and continues research in AI evaluation and reliability, all conducted independently, outside of his employment.
His most recent project, FVA‑RAG (Falsification ‑ Verification Alignment for Retrieval - Augmented Generation), applies the same scrutiny to a different challenge: ensuring that AI systems reference and verify the source material they cite. Released as a preprint on arXiv, the framework forces retrieval‑augmented models to confirm every claim against supporting documents before returning results. In benchmark tests, FVA‑RAG achieved roughly 80 % verification accuracy, outperforming leading alternatives by more than eight percentage points.
“The retrieval problem isn’t tied to any one industry,” Ravishankara explains. “Anywhere professionals rely on AI to summarize complex source material; research papers, medical records, corporate filings; there has to be a layer of verifiable grounding. That’s the gap FVA‑RAG aims to fill.”
Recognition and Service
Ravishankara’s contributions extend beyond research papers. He has completed six peer reviews for machine‑learning conference tracks, including submissions to ICLR Workshops, and served as a judge for multiple hackathons, among them Cal Hacks 12.0 at UC Berkeley. Such roles are typically reserved for recognized contributors within the research community. He also serves as a Computer Science Fellow at the Open Avenues Foundation, where he is designing and leading students through a project building production-aligned retrieval-augmented generation (RAG) systems, translating current AI research into hands-on engineering experience.
Professionally, Ravishankara works as a software engineer at Everlaw, a legal technology company based in Oakland, California, that develops software for discovery, litigation, and investigations, including AI-enabled tools. There, he has contributed across multiple product areas, including the Everlaw AI Writing Assistant, connectors for data ingestion, the Everlaw API, data-usage notifications, and core e-discovery features such as redactions. But his AI research runs in parallel, pursued independently, after hours, and on his own resources.
From Concrete to Code
Ravishankara’s roots trace back to Bengaluru, where his father built Param Machines and Moulds, a precast‑concrete manufacturing firm supplying equipment for construction and infrastructure projects across India. After early roles including a stint at DISH Network, he moved to the United States for a Master’s in Information Systems Management at Carnegie Mellon University, one of the world’s premier technology institutions. He also served as a teaching assistant in cloud computing, his first formal experience blending theory and applied systems work.
Re‑evaluating How AI Is Judged
Today, Ravishankara continues to investigate how AI models should be evaluated and what current testing methods miss.
“The field keeps building more powerful models,” he says. “But we’re still not good enough at proving they work the way we think they do. That’s the problem I keep coming back to.”
His research papers are available on IEEE Xplore and Google Scholar, and his question remains one that the entire AI community must eventually answer: does artificial intelligence truly understand what it’s doing, or merely look like it does?
